The Behavior Brief

If you hired a human intern, sat them at a desk, gave them zero instructions, and just said "be helpful," what would happen?

They would mess up. They would promise things they couldn't deliver. They would use the wrong tone with your biggest client. They would try to be "creative" with the accounting.

Yet, this is exactly how we're designing AI agents today. We slap a chat interface on a raw LLM, give it a system prompt that says "You are a helpful assistant," and pray.

As designers, we've stopped designing the soul of the software because we're too busy designing the skin.

So, we need the first new deliverable of the AI era: The Behavior Brief.

What is a Behavior Brief?

A Behavior Brief is a design artifact that defines the personality, constraints, memory, and goals of an AI agent. It's the bridge between "Product Intent" and "System Prompt."

It's not a mockup. It's a logic document.

The 4 Pillars of Behavior

1. The Goal (The North Star)

What is this agent actually trying to do? "Help the user" is not a goal. That's a platitude.

Bad: "Help the user with support."

Good: "Resolve Level 1 billing issues under $50 without escalating to a human. If the issue is over $50, gather context and escalate."

The goal should be specific enough that you could measure whether the agent achieved it.

2. The Constraints (The Electric Fence)

LLMs are people-pleasers. They will lie to make the user happy. You must design the fence.

• Financial Constraints: "Never promise a refund."
• Temporal Constraints: "Never commit to a deadline."
• Knowledge Constraints: "If you don't find the answer in the Knowledge Base, say 'I don't know.' Do not invent an answer."
• Scope Constraints: "Only discuss topics related to our product. Redirect everything else."

Write these down. Be specific. The AI will find every loophole you leave open.

3. The Tone (The Vibe)

"Friendly" is useless. "Professional" is vague. You need specific behavioral instructions.

Directive: "Use the user's first name only in the first message. Be concise (under 50 words). Do not apologize profusely; acknowledge the issue and move to the solution."

Directive: "If the user is frustrated, match their energy level but not their negativity. Acknowledge the problem, then pivot to action."

The more specific, the more consistent the behavior.

4. The Memory (The Context Window)

What does the agent know? This isn't just about data access—it's about what the agent should remember and forget.

• Session Memory: Does it remember what I said five minutes ago? (Usually yes.)
• User Memory: Does it remember I prefer dark mode from my last visit? (Depends on your architecture.)
• Entity Memory: Does it know my company's revenue from the database? (Only if you've wired it up.)
• Forgetting: What should it explicitly NOT remember? (Maybe sensitive information shared in passing.)

The Takeaway

If you can't write the Behavior Brief, you aren't ready to design the UI. Stop drawing chat bubbles. Start writing job descriptions.

The Behavior Brief is your first deliverable. Before you open Figma, open a doc.